PaddlePaddle · JiabinYang · Nov 29, 2021 · Nov 29, 2021
diff --git a/paddle/fluid/eager/auto_code_generator/eager_generator.cc b/paddle/fluid/eager/auto_code_generator/eager_generator.cc
@@ -779,7 +779,7 @@ static std::pair<std::string, std::string> GenerateForwardFunctionContents(
   ,ConstructDuplicableOutput(Out1Num)} };
 
         // According to op_proto->attrs()
-        egr::RunOp("op_type", ins, outs, attr_map,
+        egr::legacy::RunOp("op_type", ins, outs, attr_map,
   Controller.Instance().GetExpectedPlace(), {});
 
         // According to fwd_outputs_names
@@ -894,7 +894,7 @@ static std::pair<std::string, std::string> GenerateForwardFunctionContents(
   const char* FWD_TRACE_OP_TEMPLATE =
       "  paddle::framework::AttributeMap attrs = attr_map;\n"
       "  paddle::framework::AttributeMap default_attrs;\n"
-      "  egr::RunOp(\"%s\", ins, outs, attrs, \n"
+      "  egr::legacy::RunOp(\"%s\", ins, outs, attrs, \n"
       "     egr::Controller::Instance().GetExpectedPlace(),\n"
       "     &default_attrs, true, {});\n";
   std::string trace_op_str =
@@ -1052,7 +1052,7 @@ static std::string GenerateGradNodeCCContents(
     // Visit each OpBase
     for(auto iter = "grad_node->begin()"; iter < "grad_node->end()"; iter++) {
         // Simply pass entire attribute map to kernels
-        egr::RunOp("iter->Type()", ins, outs, this->attr_map_,
+        egr::legacy::RunOp("iter->Type()", ins, outs, this->attr_map_,
             egr::Controller::Instance().ExpectedPlace(), false, {});
     }
 
@@ -1180,7 +1180,7 @@ static std::string GenerateGradNodeCCContents(
         "  // Pass the entire attribute map to TraceOp\n"
         "  // The underlying kernel will pickup whatever attribute they need "
         "at runtime\n"
-        "  egr::RunOp(\"%s\", ins, outs, this->attr_map_,\n"
+        "  egr::legacy::RunOp(\"%s\", ins, outs, this->attr_map_,\n"
         "      egr::Controller::Instance().GetExpectedPlace(),\n"
         "      &this->default_attr_map_, false, {});\n";
     trace_opbase_str = paddle::string::Sprintf(TRACE_OP_TEMPLATE, op_base_type);

diff --git a/paddle/fluid/eager/legacy/amp_auto_cast.cc b/paddle/fluid/eager/legacy/amp_auto_cast.cc
@@ -20,6 +20,7 @@
 #include "paddle/fluid/framework/operator.h"
 
 namespace egr {
+namespace legacy {
 
 AmpOperators::AmpOperators()
     : allow_ops_(new std::unordered_set<std::string>()),
@@ -85,12 +86,12 @@ std::ostream& operator<<(std::ostream& os, AmpOperators& ops) {
 inline std::string GetDtypeStr(
     const std::shared_ptr<egr::EagerTensor>& tensor) {
   return paddle::framework::DataTypeToString(
-      egr::GetDtypeFromVar(tensor->Var()));
+      egr::legacy::GetDtypeFromVar(tensor->Var()));
 }
 
 inline bool NeedCast(const std::shared_ptr<egr::EagerTensor>& tensor) {
-  auto place = egr::GetPlaceFromVar(tensor->Var());
-  auto data_type = egr::GetDtypeFromVar(tensor->Var());
+  auto place = egr::legacy::GetPlaceFromVar(tensor->Var());
+  auto data_type = egr::legacy::GetDtypeFromVar(tensor->Var());
   if (paddle::platform::is_gpu_place(place) ||
       paddle::platform::is_cuda_pinned_place(place) ||
       paddle::platform::is_xpu_place(place)) {
@@ -109,7 +110,7 @@ static inline std::shared_ptr<egr::EagerTensor> CastToType(
     const std::shared_ptr<egr::EagerTensor>& tensor,
     const paddle::framework::proto::VarType::Type dst_type) {
   NameTensorMap ins = {{"X", {tensor}}};
-  auto in_data_type = egr::GetDtypeFromVar(tensor->Var());
+  auto in_data_type = egr::legacy::GetDtypeFromVar(tensor->Var());
   paddle::framework::AttributeMap attrs = {{"in_dtype", in_data_type},
                                            {"out_dtype", dst_type}};
   auto out = std::shared_ptr<egr::EagerTensor>(new egr::EagerTensor());
@@ -127,7 +128,8 @@ static inline std::shared_ptr<egr::EagerTensor> CastToType(
 static inline std::shared_ptr<egr::EagerTensor> CastToFP16(
     const std::shared_ptr<egr::EagerTensor>& tensor) {
   auto dst_type = paddle::framework::proto::VarType::FP16;
-  if (NeedCast(tensor) && (egr::GetDtypeFromVar(tensor->Var()) != dst_type)) {
+  if (NeedCast(tensor) &&
+      (egr::legacy::GetDtypeFromVar(tensor->Var()) != dst_type)) {
     return CastToType(tensor, dst_type);
   }
   return tensor;
@@ -136,7 +138,8 @@ static inline std::shared_ptr<egr::EagerTensor> CastToFP16(
 static inline std::shared_ptr<egr::EagerTensor> CastToFP32(
     const std::shared_ptr<egr::EagerTensor>& tensor) {
   auto dst_type = paddle::framework::proto::VarType::FP32;
-  if (NeedCast(tensor) && (egr::GetDtypeFromVar(tensor->Var()) != dst_type)) {
+  if (NeedCast(tensor) &&
+      (egr::legacy::GetDtypeFromVar(tensor->Var()) != dst_type)) {
     return CastToType(tensor, dst_type);
   }
   return tensor;
@@ -147,9 +150,9 @@ static inline paddle::framework::proto::VarType::Type GetPromoteType(
   auto dst_type = paddle::framework::proto::VarType::FP16;
   for (const auto& pair : ins) {
     for (const auto& tensor : pair.second) {
-      if (egr::GetDtypeFromVar(tensor->Var()) ==
+      if (egr::legacy::GetDtypeFromVar(tensor->Var()) ==
           paddle::framework::proto::VarType::FP32) {
-        dst_type = egr::GetDtypeFromVar(tensor->Var());
+        dst_type = egr::legacy::GetDtypeFromVar(tensor->Var());
         break;
       }
     }
@@ -160,7 +163,7 @@ static inline paddle::framework::proto::VarType::Type GetPromoteType(
   if (op_type == "moving_average_abs_max_scale") {
     for (const auto& pair : ins) {
       if (pair.first == "X" &&
-          egr::GetDtypeFromVar(pair.second.front()->Var()) ==
+          egr::legacy::GetDtypeFromVar(pair.second.front()->Var()) ==
               paddle::framework::proto::VarType::FP16) {
         dst_type = paddle::framework::proto::VarType::FP16;
       }
@@ -255,4 +258,5 @@ NameTensorMap CastPureFp16Inputs(const std::string& op_type,
   return new_ins;
 }
 
+}  // namespace legacy
 }  // namespace egr
diff --git a/paddle/fluid/eager/legacy/amp_auto_cast.h b/paddle/fluid/eager/legacy/amp_auto_cast.h
@@ -24,6 +24,7 @@
 #include "paddle/fluid/eager/legacy/type_def.h"
 
 namespace egr {
+namespace legacy {
 
 // NOTE(zhiqiu): only O1 and O2 are valid now
 enum class AmpLevel {
@@ -92,4 +93,5 @@ NameTensorMap AutoCastInputs(const std::string& op_type,
 NameTensorMap CastPureFp16Inputs(const std::string& op_type,
                                  const NameTensorMap& ins);
 
+}  // namespace legacy
 }  // namespace egr
diff --git a/paddle/fluid/eager/legacy/execution_context.h b/paddle/fluid/eager/legacy/execution_context.h
@@ -22,6 +22,7 @@
 #include "paddle/fluid/framework/type_defs.h"
 #include "paddle/fluid/framework/variable.h"
 namespace egr {
+namespace legacy {
 
 class EagerExecutionContext : public paddle::framework::ExecutionContext {
   using Variable = paddle::framework::Variable;
@@ -209,4 +210,5 @@ class EagerExecutionContext : public paddle::framework::ExecutionContext {
   const paddle::framework::AttributeMap& default_attrs_;
 };
 
+}  // namespace legacy
 }  // namespace egr
diff --git a/paddle/fluid/eager/legacy/infer_shape_context.h b/paddle/fluid/eager/legacy/infer_shape_context.h
@@ -25,6 +25,7 @@
 #include "paddle/fluid/framework/type_defs.h"
 #include "paddle/fluid/framework/var_type.h"
 namespace egr {
+namespace legacy {
 
 class EagerInferShapeContext : public paddle::framework::InferShapeContext {
   using DDim = paddle::framework::DDim;
@@ -401,4 +402,5 @@ class EagerInferShapeContext : public paddle::framework::InferShapeContext {
   const std::string op_type_;
 };
 
+}  // namespace legacy
 }  // namespace egr
diff --git a/paddle/fluid/eager/legacy/infer_var_type_context.h b/paddle/fluid/eager/legacy/infer_var_type_context.h
@@ -29,6 +29,7 @@
 #include "paddle/pten/include/core.h"
 
 namespace egr {
+namespace legacy {
 
 // infer var type context for imperative mode
 class TensorRuntimeInferVarTypeContext
@@ -255,4 +256,5 @@ class TensorRuntimeInferVarTypeContext
   const paddle::framework::AttributeMap& default_attrs_;
 };
 
+}  // namespace legacy
 }  // namespace egr
diff --git a/paddle/fluid/eager/legacy/op_runner.cc b/paddle/fluid/eager/legacy/op_runner.cc
@@ -30,6 +30,7 @@ DECLARE_string(tracer_mkldnn_ops_on);
 DECLARE_string(tracer_mkldnn_ops_off);
 
 namespace egr {
+namespace legacy {
 
 void OpRunImpl(const paddle::framework::OperatorBase& op,
                const NameTensorMap& ins, const NameTensorMap& outs,
@@ -43,8 +44,8 @@ void OpRunImpl(const paddle::framework::OperatorBase& op,
                      "Only support operator with kernel in Dygraph mode."));
   auto& info = op.Info();
   if (info.infer_var_type_) {
-    egr::TensorRuntimeInferVarTypeContext infer_var_type_ctx(ins, outs, attrs,
-                                                             default_attrs);
+    egr::legacy::TensorRuntimeInferVarTypeContext infer_var_type_ctx(
+        ins, outs, attrs, default_attrs);
     info.infer_var_type_(&infer_var_type_ctx);
   }
 
@@ -76,10 +77,10 @@ void OpRunImpl(const paddle::framework::OperatorBase& op,
    * after the execution of op, but the original input is directly
    * overwritten in the previous dynamic graph implemention.
    */
-  auto prepared_op = egr::PreparedOp::Prepare(ins, outs, *op_kernel, place,
-                                              attrs, default_attrs);
+  auto prepared_op = egr::legacy::PreparedOp::Prepare(
+      ins, outs, *op_kernel, place, attrs, default_attrs);
   auto tmp_ins_ptr =
-      egr::PrepareData(*op_kernel, ins, prepared_op.kernel_type());
+      egr::legacy::PrepareData(*op_kernel, ins, prepared_op.kernel_type());
   if (tmp_ins_ptr == nullptr) {
     prepared_op.Run(ins, outs, attrs, default_attrs);
   } else {
@@ -188,4 +189,6 @@ void RunOp(const std::string& type, const NameTensorMap& ins,
   //   program_desc_tracer_->InsertOp(type, new_ins, outs, attrs);
   // }
 }
+
+}  // namespace legacy
 }  // namespace egr
diff --git a/paddle/fluid/eager/legacy/op_runner.h b/paddle/fluid/eager/legacy/op_runner.h
@@ -14,16 +14,18 @@
 
 #pragma once
 #include "paddle/fluid/eager/legacy/type_def.h"
-// TODO(Jiabin): We should not depends on this header remove it later
 #include "paddle/fluid/imperative/jit/program_desc_tracer.h"
 #include "paddle/pten/core/tensor_meta.h"
 
 namespace egr {
+namespace legacy {
 
 void RunOp(const std::string& type, const NameTensorMap& ins,
            const NameTensorMap& outs, paddle::framework::AttributeMap attrs,
            const paddle::platform::Place& place,
            paddle::framework::AttributeMap* default_attrs,
            bool override_default_attr_map,
            const std::map<std::string, std::string>& inplace_map = {});
-}
+
+}  // namespace legacy
+}  // namespace egr
diff --git a/paddle/fluid/eager/legacy/prepared_operator.cc b/paddle/fluid/eager/legacy/prepared_operator.cc
@@ -26,6 +26,7 @@ DECLARE_bool(check_nan_inf);
 DECLARE_bool(run_pten_kernel);
 
 namespace egr {
+namespace legacy {
 
 const paddle::framework::Tensor* GetTensorFromVar(
     const paddle::framework::Variable& var) {
@@ -96,9 +97,9 @@ PreparedOp PrepareImpl(const NameTensorMap& ins, const NameTensorMap& outs,
 #endif
 
   // 1. get expected kernel key
-  auto dygraph_exe_ctx =
-      egr::EagerExecutionContext(op, paddle::framework::Scope(), *dev_ctx, ctx,
-                                 ins, outs, attrs, default_attrs);
+  auto dygraph_exe_ctx = egr::legacy::EagerExecutionContext(
+      op, paddle::framework::Scope(), *dev_ctx, ctx, ins, outs, attrs,
+      default_attrs);
   auto expected_kernel_key = op.GetExpectedKernelType(dygraph_exe_ctx);
   VLOG(3) << "expected_kernel_key:" << expected_kernel_key;
 
@@ -251,4 +252,6 @@ std::shared_ptr<NameTensorMap> PrepareData(
   }
   return tmp_ins_ptr;
 }
+
+}  // namespace legacy
 }  // namespace egr
diff --git a/paddle/fluid/eager/legacy/prepared_operator.h b/paddle/fluid/eager/legacy/prepared_operator.h
@@ -40,6 +40,7 @@ class DeviceContext;
 }  // namespace paddle
 
 namespace egr {
+namespace legacy {
 
 const paddle::framework::Tensor* GetTensorFromVar(
     const paddle::framework::Variable& var);
@@ -79,4 +80,5 @@ class PreparedOp {
   paddle::platform::DeviceContext* dev_ctx_;
 };
 
+}  // namespace legacy
 }  // namespace egr
diff --git a/paddle/fluid/eager/legacy/tensor_helper.cc b/paddle/fluid/eager/legacy/tensor_helper.cc
@@ -25,6 +25,7 @@
 #include "paddle/fluid/platform/place.h"
 
 namespace egr {
+namespace legacy {
 
 void InitializeVariable(paddle::framework::Variable *var,
                         paddle::framework::proto::VarType::Type var_type) {
@@ -108,4 +109,6 @@ const paddle::platform::Place &GetPlaceFromVar(
         paddle::framework::ToTypeName(var.Type())));
   }
 }
+
+}  // namespace legacy
 }  // namespace egr
diff --git a/paddle/fluid/eager/legacy/tensor_helper.h b/paddle/fluid/eager/legacy/tensor_helper.h
@@ -19,6 +19,8 @@
 #include "paddle/pten/api/all.h"
 #include "paddle/pten/include/core.h"
 namespace egr {
+namespace legacy {
+
 void InitializeVariable(paddle::framework::Variable* var,
                         paddle::framework::proto::VarType::Type var_type);
 paddle::framework::proto::VarType::Type GetDtypeFromVar(
@@ -27,4 +29,6 @@ const paddle::platform::Place& GetPlaceFromVar(
     const paddle::framework::Variable& var);
 void CopyVariable(const paddle::framework::Variable& src_var,
                   paddle::framework::Variable* dst_var);
-}
+
+}  // namespace legacy
+}  // namespace egr
diff --git a/paddle/fluid/eager/legacy/type_def.h b/paddle/fluid/eager/legacy/type_def.h
@@ -22,6 +22,9 @@
 namespace egr {
 
 class EagerTensor;
+
+namespace legacy {
+
 namespace details {
 template <typename T>
 struct NameVarMapTrait {};
@@ -36,4 +39,6 @@ template <typename T>
 using NameMap = typename details::NameVarMapTrait<T>::Type;
 
 using NameTensorMap = NameMap<EagerTensor>;
+
+}  // namespace legacy
 }  // namespace egr
diff --git a/paddle/fluid/framework/details/nan_inf_utils.h b/paddle/fluid/framework/details/nan_inf_utils.h
@@ -55,9 +55,9 @@ void CheckOpHasNanOrInfInDygraph(const std::string& op_type,
 }
 
 template <typename TensorType>
-static void CheckOpHasNanOrInfInEager(const std::string& op_type,
-                                      const egr::NameMap<TensorType>& op_outs,
-                                      platform::Place place) {
+static void CheckOpHasNanOrInfInEager(
+    const std::string& op_type, const egr::legacy::NameMap<TensorType>& op_outs,
+    platform::Place place) {
   for (const auto& pair : op_outs) {
     for (const auto& tensor : pair.second) {
       auto* var = tensor->MutableVar();